More and more universities and corporations are starting to provide videotaped lectures online for knowledge sharing and learning. Segmenting lecture videos into short clips by topic can extract the hidden information structure of the videos and facilitate information searching and learning. Manual segmentation has high accuracy rates but is very labor intensive. In order to develop a high performance automated segmentation method for lecture videos, we conducted a case study to learn the segmentation process of humans and the effective segmentation features used in the process. Based on the findings from the case study, we designed an automated segmentation approach with two phases: initial segmentation and segmentation refinement. The approach combines segmentation features from three information sources of video (speech text transcript, audio and video) and makes use of various knowledge sources such as world knowledge and domain knowledge. Our preliminary results show that the proposed two-phase approach is promising.